Abstract
In this letter, we propose a novel three-layer convex network termed as 3CN for domain adaptation in multitemporal very high resolution (VHR) remote sensing images. 3CN is composed of three main layers: 1) mapping source training samples to the target domain via a special single-layer feedforward neural network called extreme learning machine (ELM); 2) target image classification via ELM too; and 3) spatial regularization via the random-walker algorithm, which models the target image as a lattice graph and then minimizes an energy functional. This network is convex because all three layers have closed-form solutions. In the preprocessing step, we use scale-invariant feature transform to extract a set of matching key points called inliers from source and target images. Then, these inliers are used by layer 1 of 3CN to spectrally map the source training samples to the target domain. Next, in layer 2, we use the mapped training set to classify the target image. In layer 3, we exploit the spatial contextual information in the target image to reduce noise and generate an improved classification map. In the final step, we iteratively fine-tune the network to increase its discrimination ability and reduce the shift between the target and source domains. In the experiments, we report and discuss the results of the proposed method on two data sets of VHR image pairs acquired by IKONOS-2 and GeoEye-1.